1 /* 2 * An async IO implementation for Linux 3 * Written by Benjamin LaHaise <bcrl@kvack.org> 4 * 5 * Implements an efficient asynchronous io interface. 6 * 7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved. 8 * 9 * See ../COPYING for licensing terms. 10 */ 11 #include <linux/kernel.h> 12 #include <linux/init.h> 13 #include <linux/errno.h> 14 #include <linux/time.h> 15 #include <linux/aio_abi.h> 16 #include <linux/module.h> 17 #include <linux/syscalls.h> 18 #include <linux/uio.h> 19 20 #define DEBUG 0 21 22 #include <linux/sched.h> 23 #include <linux/fs.h> 24 #include <linux/file.h> 25 #include <linux/mm.h> 26 #include <linux/mman.h> 27 #include <linux/slab.h> 28 #include <linux/timer.h> 29 #include <linux/aio.h> 30 #include <linux/highmem.h> 31 #include <linux/workqueue.h> 32 #include <linux/security.h> 33 34 #include <asm/kmap_types.h> 35 #include <asm/uaccess.h> 36 #include <asm/mmu_context.h> 37 38 #if DEBUG > 1 39 #define dprintk printk 40 #else 41 #define dprintk(x...) do { ; } while (0) 42 #endif 43 44 /*------ sysctl variables----*/ 45 static DEFINE_SPINLOCK(aio_nr_lock); 46 unsigned long aio_nr; /* current system wide number of aio requests */ 47 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 48 /*----end sysctl variables---*/ 49 50 static struct kmem_cache *kiocb_cachep; 51 static struct kmem_cache *kioctx_cachep; 52 53 static struct workqueue_struct *aio_wq; 54 55 /* Used for rare fput completion. */ 56 static void aio_fput_routine(struct work_struct *); 57 static DECLARE_WORK(fput_work, aio_fput_routine); 58 59 static DEFINE_SPINLOCK(fput_lock); 60 static LIST_HEAD(fput_head); 61 62 static void aio_kick_handler(struct work_struct *); 63 static void aio_queue_work(struct kioctx *); 64 65 /* aio_setup 66 * Creates the slab caches used by the aio routines, panic on 67 * failure as this is done early during the boot sequence. 68 */ 69 static int __init aio_setup(void) 70 { 71 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb), 72 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL); 73 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx), 74 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL); 75 76 aio_wq = create_workqueue("aio"); 77 78 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page)); 79 80 return 0; 81 } 82 83 static void aio_free_ring(struct kioctx *ctx) 84 { 85 struct aio_ring_info *info = &ctx->ring_info; 86 long i; 87 88 for (i=0; i<info->nr_pages; i++) 89 put_page(info->ring_pages[i]); 90 91 if (info->mmap_size) { 92 down_write(&ctx->mm->mmap_sem); 93 do_munmap(ctx->mm, info->mmap_base, info->mmap_size); 94 up_write(&ctx->mm->mmap_sem); 95 } 96 97 if (info->ring_pages && info->ring_pages != info->internal_pages) 98 kfree(info->ring_pages); 99 info->ring_pages = NULL; 100 info->nr = 0; 101 } 102 103 static int aio_setup_ring(struct kioctx *ctx) 104 { 105 struct aio_ring *ring; 106 struct aio_ring_info *info = &ctx->ring_info; 107 unsigned nr_events = ctx->max_reqs; 108 unsigned long size; 109 int nr_pages; 110 111 /* Compensate for the ring buffer's head/tail overlap entry */ 112 nr_events += 2; /* 1 is required, 2 for good luck */ 113 114 size = sizeof(struct aio_ring); 115 size += sizeof(struct io_event) * nr_events; 116 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT; 117 118 if (nr_pages < 0) 119 return -EINVAL; 120 121 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event); 122 123 info->nr = 0; 124 info->ring_pages = info->internal_pages; 125 if (nr_pages > AIO_RING_PAGES) { 126 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); 127 if (!info->ring_pages) 128 return -ENOMEM; 129 } 130 131 info->mmap_size = nr_pages * PAGE_SIZE; 132 dprintk("attempting mmap of %lu bytes\n", info->mmap_size); 133 down_write(&ctx->mm->mmap_sem); 134 info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 135 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, 136 0); 137 if (IS_ERR((void *)info->mmap_base)) { 138 up_write(&ctx->mm->mmap_sem); 139 printk("mmap err: %ld\n", -info->mmap_base); 140 info->mmap_size = 0; 141 aio_free_ring(ctx); 142 return -EAGAIN; 143 } 144 145 dprintk("mmap address: 0x%08lx\n", info->mmap_base); 146 info->nr_pages = get_user_pages(current, ctx->mm, 147 info->mmap_base, nr_pages, 148 1, 0, info->ring_pages, NULL); 149 up_write(&ctx->mm->mmap_sem); 150 151 if (unlikely(info->nr_pages != nr_pages)) { 152 aio_free_ring(ctx); 153 return -EAGAIN; 154 } 155 156 ctx->user_id = info->mmap_base; 157 158 info->nr = nr_events; /* trusted copy */ 159 160 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 161 ring->nr = nr_events; /* user copy */ 162 ring->id = ctx->user_id; 163 ring->head = ring->tail = 0; 164 ring->magic = AIO_RING_MAGIC; 165 ring->compat_features = AIO_RING_COMPAT_FEATURES; 166 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 167 ring->header_length = sizeof(struct aio_ring); 168 kunmap_atomic(ring, KM_USER0); 169 170 return 0; 171 } 172 173 174 /* aio_ring_event: returns a pointer to the event at the given index from 175 * kmap_atomic(, km). Release the pointer with put_aio_ring_event(); 176 */ 177 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 178 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 179 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 180 181 #define aio_ring_event(info, nr, km) ({ \ 182 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \ 183 struct io_event *__event; \ 184 __event = kmap_atomic( \ 185 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \ 186 __event += pos % AIO_EVENTS_PER_PAGE; \ 187 __event; \ 188 }) 189 190 #define put_aio_ring_event(event, km) do { \ 191 struct io_event *__event = (event); \ 192 (void)__event; \ 193 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \ 194 } while(0) 195 196 /* ioctx_alloc 197 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 198 */ 199 static struct kioctx *ioctx_alloc(unsigned nr_events) 200 { 201 struct mm_struct *mm; 202 struct kioctx *ctx; 203 204 /* Prevent overflows */ 205 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 206 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 207 pr_debug("ENOMEM: nr_events too high\n"); 208 return ERR_PTR(-EINVAL); 209 } 210 211 if ((unsigned long)nr_events > aio_max_nr) 212 return ERR_PTR(-EAGAIN); 213 214 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL); 215 if (!ctx) 216 return ERR_PTR(-ENOMEM); 217 218 memset(ctx, 0, sizeof(*ctx)); 219 ctx->max_reqs = nr_events; 220 mm = ctx->mm = current->mm; 221 atomic_inc(&mm->mm_count); 222 223 atomic_set(&ctx->users, 1); 224 spin_lock_init(&ctx->ctx_lock); 225 spin_lock_init(&ctx->ring_info.ring_lock); 226 init_waitqueue_head(&ctx->wait); 227 228 INIT_LIST_HEAD(&ctx->active_reqs); 229 INIT_LIST_HEAD(&ctx->run_list); 230 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler); 231 232 if (aio_setup_ring(ctx) < 0) 233 goto out_freectx; 234 235 /* limit the number of system wide aios */ 236 spin_lock(&aio_nr_lock); 237 if (aio_nr + ctx->max_reqs > aio_max_nr || 238 aio_nr + ctx->max_reqs < aio_nr) 239 ctx->max_reqs = 0; 240 else 241 aio_nr += ctx->max_reqs; 242 spin_unlock(&aio_nr_lock); 243 if (ctx->max_reqs == 0) 244 goto out_cleanup; 245 246 /* now link into global list. kludge. FIXME */ 247 write_lock(&mm->ioctx_list_lock); 248 ctx->next = mm->ioctx_list; 249 mm->ioctx_list = ctx; 250 write_unlock(&mm->ioctx_list_lock); 251 252 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 253 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 254 return ctx; 255 256 out_cleanup: 257 __put_ioctx(ctx); 258 return ERR_PTR(-EAGAIN); 259 260 out_freectx: 261 mmdrop(mm); 262 kmem_cache_free(kioctx_cachep, ctx); 263 ctx = ERR_PTR(-ENOMEM); 264 265 dprintk("aio: error allocating ioctx %p\n", ctx); 266 return ctx; 267 } 268 269 /* aio_cancel_all 270 * Cancels all outstanding aio requests on an aio context. Used 271 * when the processes owning a context have all exited to encourage 272 * the rapid destruction of the kioctx. 273 */ 274 static void aio_cancel_all(struct kioctx *ctx) 275 { 276 int (*cancel)(struct kiocb *, struct io_event *); 277 struct io_event res; 278 spin_lock_irq(&ctx->ctx_lock); 279 ctx->dead = 1; 280 while (!list_empty(&ctx->active_reqs)) { 281 struct list_head *pos = ctx->active_reqs.next; 282 struct kiocb *iocb = list_kiocb(pos); 283 list_del_init(&iocb->ki_list); 284 cancel = iocb->ki_cancel; 285 kiocbSetCancelled(iocb); 286 if (cancel) { 287 iocb->ki_users++; 288 spin_unlock_irq(&ctx->ctx_lock); 289 cancel(iocb, &res); 290 spin_lock_irq(&ctx->ctx_lock); 291 } 292 } 293 spin_unlock_irq(&ctx->ctx_lock); 294 } 295 296 static void wait_for_all_aios(struct kioctx *ctx) 297 { 298 struct task_struct *tsk = current; 299 DECLARE_WAITQUEUE(wait, tsk); 300 301 spin_lock_irq(&ctx->ctx_lock); 302 if (!ctx->reqs_active) 303 goto out; 304 305 add_wait_queue(&ctx->wait, &wait); 306 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 307 while (ctx->reqs_active) { 308 spin_unlock_irq(&ctx->ctx_lock); 309 schedule(); 310 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 311 spin_lock_irq(&ctx->ctx_lock); 312 } 313 __set_task_state(tsk, TASK_RUNNING); 314 remove_wait_queue(&ctx->wait, &wait); 315 316 out: 317 spin_unlock_irq(&ctx->ctx_lock); 318 } 319 320 /* wait_on_sync_kiocb: 321 * Waits on the given sync kiocb to complete. 322 */ 323 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb) 324 { 325 while (iocb->ki_users) { 326 set_current_state(TASK_UNINTERRUPTIBLE); 327 if (!iocb->ki_users) 328 break; 329 schedule(); 330 } 331 __set_current_state(TASK_RUNNING); 332 return iocb->ki_user_data; 333 } 334 335 /* exit_aio: called when the last user of mm goes away. At this point, 336 * there is no way for any new requests to be submited or any of the 337 * io_* syscalls to be called on the context. However, there may be 338 * outstanding requests which hold references to the context; as they 339 * go away, they will call put_ioctx and release any pinned memory 340 * associated with the request (held via struct page * references). 341 */ 342 void fastcall exit_aio(struct mm_struct *mm) 343 { 344 struct kioctx *ctx = mm->ioctx_list; 345 mm->ioctx_list = NULL; 346 while (ctx) { 347 struct kioctx *next = ctx->next; 348 ctx->next = NULL; 349 aio_cancel_all(ctx); 350 351 wait_for_all_aios(ctx); 352 /* 353 * this is an overkill, but ensures we don't leave 354 * the ctx on the aio_wq 355 */ 356 flush_workqueue(aio_wq); 357 358 if (1 != atomic_read(&ctx->users)) 359 printk(KERN_DEBUG 360 "exit_aio:ioctx still alive: %d %d %d\n", 361 atomic_read(&ctx->users), ctx->dead, 362 ctx->reqs_active); 363 put_ioctx(ctx); 364 ctx = next; 365 } 366 } 367 368 /* __put_ioctx 369 * Called when the last user of an aio context has gone away, 370 * and the struct needs to be freed. 371 */ 372 void fastcall __put_ioctx(struct kioctx *ctx) 373 { 374 unsigned nr_events = ctx->max_reqs; 375 376 BUG_ON(ctx->reqs_active); 377 378 cancel_delayed_work(&ctx->wq); 379 flush_workqueue(aio_wq); 380 aio_free_ring(ctx); 381 mmdrop(ctx->mm); 382 ctx->mm = NULL; 383 pr_debug("__put_ioctx: freeing %p\n", ctx); 384 kmem_cache_free(kioctx_cachep, ctx); 385 386 if (nr_events) { 387 spin_lock(&aio_nr_lock); 388 BUG_ON(aio_nr - nr_events > aio_nr); 389 aio_nr -= nr_events; 390 spin_unlock(&aio_nr_lock); 391 } 392 } 393 394 /* aio_get_req 395 * Allocate a slot for an aio request. Increments the users count 396 * of the kioctx so that the kioctx stays around until all requests are 397 * complete. Returns NULL if no requests are free. 398 * 399 * Returns with kiocb->users set to 2. The io submit code path holds 400 * an extra reference while submitting the i/o. 401 * This prevents races between the aio code path referencing the 402 * req (after submitting it) and aio_complete() freeing the req. 403 */ 404 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx)); 405 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx) 406 { 407 struct kiocb *req = NULL; 408 struct aio_ring *ring; 409 int okay = 0; 410 411 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 412 if (unlikely(!req)) 413 return NULL; 414 415 req->ki_flags = 0; 416 req->ki_users = 2; 417 req->ki_key = 0; 418 req->ki_ctx = ctx; 419 req->ki_cancel = NULL; 420 req->ki_retry = NULL; 421 req->ki_dtor = NULL; 422 req->private = NULL; 423 req->ki_iovec = NULL; 424 INIT_LIST_HEAD(&req->ki_run_list); 425 426 /* Check if the completion queue has enough free space to 427 * accept an event from this io. 428 */ 429 spin_lock_irq(&ctx->ctx_lock); 430 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 431 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 432 list_add(&req->ki_list, &ctx->active_reqs); 433 ctx->reqs_active++; 434 okay = 1; 435 } 436 kunmap_atomic(ring, KM_USER0); 437 spin_unlock_irq(&ctx->ctx_lock); 438 439 if (!okay) { 440 kmem_cache_free(kiocb_cachep, req); 441 req = NULL; 442 } 443 444 return req; 445 } 446 447 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 448 { 449 struct kiocb *req; 450 /* Handle a potential starvation case -- should be exceedingly rare as 451 * requests will be stuck on fput_head only if the aio_fput_routine is 452 * delayed and the requests were the last user of the struct file. 453 */ 454 req = __aio_get_req(ctx); 455 if (unlikely(NULL == req)) { 456 aio_fput_routine(NULL); 457 req = __aio_get_req(ctx); 458 } 459 return req; 460 } 461 462 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 463 { 464 assert_spin_locked(&ctx->ctx_lock); 465 466 if (req->ki_dtor) 467 req->ki_dtor(req); 468 if (req->ki_iovec != &req->ki_inline_vec) 469 kfree(req->ki_iovec); 470 kmem_cache_free(kiocb_cachep, req); 471 ctx->reqs_active--; 472 473 if (unlikely(!ctx->reqs_active && ctx->dead)) 474 wake_up(&ctx->wait); 475 } 476 477 static void aio_fput_routine(struct work_struct *data) 478 { 479 spin_lock_irq(&fput_lock); 480 while (likely(!list_empty(&fput_head))) { 481 struct kiocb *req = list_kiocb(fput_head.next); 482 struct kioctx *ctx = req->ki_ctx; 483 484 list_del(&req->ki_list); 485 spin_unlock_irq(&fput_lock); 486 487 /* Complete the fput */ 488 __fput(req->ki_filp); 489 490 /* Link the iocb into the context's free list */ 491 spin_lock_irq(&ctx->ctx_lock); 492 really_put_req(ctx, req); 493 spin_unlock_irq(&ctx->ctx_lock); 494 495 put_ioctx(ctx); 496 spin_lock_irq(&fput_lock); 497 } 498 spin_unlock_irq(&fput_lock); 499 } 500 501 /* __aio_put_req 502 * Returns true if this put was the last user of the request. 503 */ 504 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 505 { 506 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n", 507 req, atomic_read(&req->ki_filp->f_count)); 508 509 assert_spin_locked(&ctx->ctx_lock); 510 511 req->ki_users --; 512 BUG_ON(req->ki_users < 0); 513 if (likely(req->ki_users)) 514 return 0; 515 list_del(&req->ki_list); /* remove from active_reqs */ 516 req->ki_cancel = NULL; 517 req->ki_retry = NULL; 518 519 /* Must be done under the lock to serialise against cancellation. 520 * Call this aio_fput as it duplicates fput via the fput_work. 521 */ 522 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) { 523 get_ioctx(ctx); 524 spin_lock(&fput_lock); 525 list_add(&req->ki_list, &fput_head); 526 spin_unlock(&fput_lock); 527 queue_work(aio_wq, &fput_work); 528 } else 529 really_put_req(ctx, req); 530 return 1; 531 } 532 533 /* aio_put_req 534 * Returns true if this put was the last user of the kiocb, 535 * false if the request is still in use. 536 */ 537 int fastcall aio_put_req(struct kiocb *req) 538 { 539 struct kioctx *ctx = req->ki_ctx; 540 int ret; 541 spin_lock_irq(&ctx->ctx_lock); 542 ret = __aio_put_req(ctx, req); 543 spin_unlock_irq(&ctx->ctx_lock); 544 return ret; 545 } 546 547 /* Lookup an ioctx id. ioctx_list is lockless for reads. 548 * FIXME: this is O(n) and is only suitable for development. 549 */ 550 struct kioctx *lookup_ioctx(unsigned long ctx_id) 551 { 552 struct kioctx *ioctx; 553 struct mm_struct *mm; 554 555 mm = current->mm; 556 read_lock(&mm->ioctx_list_lock); 557 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next) 558 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) { 559 get_ioctx(ioctx); 560 break; 561 } 562 read_unlock(&mm->ioctx_list_lock); 563 564 return ioctx; 565 } 566 567 /* 568 * use_mm 569 * Makes the calling kernel thread take on the specified 570 * mm context. 571 * Called by the retry thread execute retries within the 572 * iocb issuer's mm context, so that copy_from/to_user 573 * operations work seamlessly for aio. 574 * (Note: this routine is intended to be called only 575 * from a kernel thread context) 576 */ 577 static void use_mm(struct mm_struct *mm) 578 { 579 struct mm_struct *active_mm; 580 struct task_struct *tsk = current; 581 582 task_lock(tsk); 583 tsk->flags |= PF_BORROWED_MM; 584 active_mm = tsk->active_mm; 585 atomic_inc(&mm->mm_count); 586 tsk->mm = mm; 587 tsk->active_mm = mm; 588 /* 589 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise 590 * it won't work. Update it accordingly if you change it here 591 */ 592 switch_mm(active_mm, mm, tsk); 593 task_unlock(tsk); 594 595 mmdrop(active_mm); 596 } 597 598 /* 599 * unuse_mm 600 * Reverses the effect of use_mm, i.e. releases the 601 * specified mm context which was earlier taken on 602 * by the calling kernel thread 603 * (Note: this routine is intended to be called only 604 * from a kernel thread context) 605 */ 606 static void unuse_mm(struct mm_struct *mm) 607 { 608 struct task_struct *tsk = current; 609 610 task_lock(tsk); 611 tsk->flags &= ~PF_BORROWED_MM; 612 tsk->mm = NULL; 613 /* active_mm is still 'mm' */ 614 enter_lazy_tlb(mm, tsk); 615 task_unlock(tsk); 616 } 617 618 /* 619 * Queue up a kiocb to be retried. Assumes that the kiocb 620 * has already been marked as kicked, and places it on 621 * the retry run list for the corresponding ioctx, if it 622 * isn't already queued. Returns 1 if it actually queued 623 * the kiocb (to tell the caller to activate the work 624 * queue to process it), or 0, if it found that it was 625 * already queued. 626 */ 627 static inline int __queue_kicked_iocb(struct kiocb *iocb) 628 { 629 struct kioctx *ctx = iocb->ki_ctx; 630 631 assert_spin_locked(&ctx->ctx_lock); 632 633 if (list_empty(&iocb->ki_run_list)) { 634 list_add_tail(&iocb->ki_run_list, 635 &ctx->run_list); 636 return 1; 637 } 638 return 0; 639 } 640 641 /* aio_run_iocb 642 * This is the core aio execution routine. It is 643 * invoked both for initial i/o submission and 644 * subsequent retries via the aio_kick_handler. 645 * Expects to be invoked with iocb->ki_ctx->lock 646 * already held. The lock is released and reacquired 647 * as needed during processing. 648 * 649 * Calls the iocb retry method (already setup for the 650 * iocb on initial submission) for operation specific 651 * handling, but takes care of most of common retry 652 * execution details for a given iocb. The retry method 653 * needs to be non-blocking as far as possible, to avoid 654 * holding up other iocbs waiting to be serviced by the 655 * retry kernel thread. 656 * 657 * The trickier parts in this code have to do with 658 * ensuring that only one retry instance is in progress 659 * for a given iocb at any time. Providing that guarantee 660 * simplifies the coding of individual aio operations as 661 * it avoids various potential races. 662 */ 663 static ssize_t aio_run_iocb(struct kiocb *iocb) 664 { 665 struct kioctx *ctx = iocb->ki_ctx; 666 ssize_t (*retry)(struct kiocb *); 667 ssize_t ret; 668 669 if (!(retry = iocb->ki_retry)) { 670 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 671 return 0; 672 } 673 674 /* 675 * We don't want the next retry iteration for this 676 * operation to start until this one has returned and 677 * updated the iocb state. However, wait_queue functions 678 * can trigger a kick_iocb from interrupt context in the 679 * meantime, indicating that data is available for the next 680 * iteration. We want to remember that and enable the 681 * next retry iteration _after_ we are through with 682 * this one. 683 * 684 * So, in order to be able to register a "kick", but 685 * prevent it from being queued now, we clear the kick 686 * flag, but make the kick code *think* that the iocb is 687 * still on the run list until we are actually done. 688 * When we are done with this iteration, we check if 689 * the iocb was kicked in the meantime and if so, queue 690 * it up afresh. 691 */ 692 693 kiocbClearKicked(iocb); 694 695 /* 696 * This is so that aio_complete knows it doesn't need to 697 * pull the iocb off the run list (We can't just call 698 * INIT_LIST_HEAD because we don't want a kick_iocb to 699 * queue this on the run list yet) 700 */ 701 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 702 spin_unlock_irq(&ctx->ctx_lock); 703 704 /* Quit retrying if the i/o has been cancelled */ 705 if (kiocbIsCancelled(iocb)) { 706 ret = -EINTR; 707 aio_complete(iocb, ret, 0); 708 /* must not access the iocb after this */ 709 goto out; 710 } 711 712 /* 713 * Now we are all set to call the retry method in async 714 * context. By setting this thread's io_wait context 715 * to point to the wait queue entry inside the currently 716 * running iocb for the duration of the retry, we ensure 717 * that async notification wakeups are queued by the 718 * operation instead of blocking waits, and when notified, 719 * cause the iocb to be kicked for continuation (through 720 * the aio_wake_function callback). 721 */ 722 BUG_ON(current->io_wait != NULL); 723 current->io_wait = &iocb->ki_wait; 724 ret = retry(iocb); 725 current->io_wait = NULL; 726 727 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 728 BUG_ON(!list_empty(&iocb->ki_wait.task_list)); 729 aio_complete(iocb, ret, 0); 730 } 731 out: 732 spin_lock_irq(&ctx->ctx_lock); 733 734 if (-EIOCBRETRY == ret) { 735 /* 736 * OK, now that we are done with this iteration 737 * and know that there is more left to go, 738 * this is where we let go so that a subsequent 739 * "kick" can start the next iteration 740 */ 741 742 /* will make __queue_kicked_iocb succeed from here on */ 743 INIT_LIST_HEAD(&iocb->ki_run_list); 744 /* we must queue the next iteration ourselves, if it 745 * has already been kicked */ 746 if (kiocbIsKicked(iocb)) { 747 __queue_kicked_iocb(iocb); 748 749 /* 750 * __queue_kicked_iocb will always return 1 here, because 751 * iocb->ki_run_list is empty at this point so it should 752 * be safe to unconditionally queue the context into the 753 * work queue. 754 */ 755 aio_queue_work(ctx); 756 } 757 } 758 return ret; 759 } 760 761 /* 762 * __aio_run_iocbs: 763 * Process all pending retries queued on the ioctx 764 * run list. 765 * Assumes it is operating within the aio issuer's mm 766 * context. 767 */ 768 static int __aio_run_iocbs(struct kioctx *ctx) 769 { 770 struct kiocb *iocb; 771 struct list_head run_list; 772 773 assert_spin_locked(&ctx->ctx_lock); 774 775 list_replace_init(&ctx->run_list, &run_list); 776 while (!list_empty(&run_list)) { 777 iocb = list_entry(run_list.next, struct kiocb, 778 ki_run_list); 779 list_del(&iocb->ki_run_list); 780 /* 781 * Hold an extra reference while retrying i/o. 782 */ 783 iocb->ki_users++; /* grab extra reference */ 784 aio_run_iocb(iocb); 785 __aio_put_req(ctx, iocb); 786 } 787 if (!list_empty(&ctx->run_list)) 788 return 1; 789 return 0; 790 } 791 792 static void aio_queue_work(struct kioctx * ctx) 793 { 794 unsigned long timeout; 795 /* 796 * if someone is waiting, get the work started right 797 * away, otherwise, use a longer delay 798 */ 799 smp_mb(); 800 if (waitqueue_active(&ctx->wait)) 801 timeout = 1; 802 else 803 timeout = HZ/10; 804 queue_delayed_work(aio_wq, &ctx->wq, timeout); 805 } 806 807 808 /* 809 * aio_run_iocbs: 810 * Process all pending retries queued on the ioctx 811 * run list. 812 * Assumes it is operating within the aio issuer's mm 813 * context. 814 */ 815 static inline void aio_run_iocbs(struct kioctx *ctx) 816 { 817 int requeue; 818 819 spin_lock_irq(&ctx->ctx_lock); 820 821 requeue = __aio_run_iocbs(ctx); 822 spin_unlock_irq(&ctx->ctx_lock); 823 if (requeue) 824 aio_queue_work(ctx); 825 } 826 827 /* 828 * just like aio_run_iocbs, but keeps running them until 829 * the list stays empty 830 */ 831 static inline void aio_run_all_iocbs(struct kioctx *ctx) 832 { 833 spin_lock_irq(&ctx->ctx_lock); 834 while (__aio_run_iocbs(ctx)) 835 ; 836 spin_unlock_irq(&ctx->ctx_lock); 837 } 838 839 /* 840 * aio_kick_handler: 841 * Work queue handler triggered to process pending 842 * retries on an ioctx. Takes on the aio issuer's 843 * mm context before running the iocbs, so that 844 * copy_xxx_user operates on the issuer's address 845 * space. 846 * Run on aiod's context. 847 */ 848 static void aio_kick_handler(struct work_struct *work) 849 { 850 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 851 mm_segment_t oldfs = get_fs(); 852 struct mm_struct *mm; 853 int requeue; 854 855 set_fs(USER_DS); 856 use_mm(ctx->mm); 857 spin_lock_irq(&ctx->ctx_lock); 858 requeue =__aio_run_iocbs(ctx); 859 mm = ctx->mm; 860 spin_unlock_irq(&ctx->ctx_lock); 861 unuse_mm(mm); 862 set_fs(oldfs); 863 /* 864 * we're in a worker thread already, don't use queue_delayed_work, 865 */ 866 if (requeue) 867 queue_delayed_work(aio_wq, &ctx->wq, 0); 868 } 869 870 871 /* 872 * Called by kick_iocb to queue the kiocb for retry 873 * and if required activate the aio work queue to process 874 * it 875 */ 876 static void try_queue_kicked_iocb(struct kiocb *iocb) 877 { 878 struct kioctx *ctx = iocb->ki_ctx; 879 unsigned long flags; 880 int run = 0; 881 882 /* We're supposed to be the only path putting the iocb back on the run 883 * list. If we find that the iocb is *back* on a wait queue already 884 * than retry has happened before we could queue the iocb. This also 885 * means that the retry could have completed and freed our iocb, no 886 * good. */ 887 BUG_ON((!list_empty(&iocb->ki_wait.task_list))); 888 889 spin_lock_irqsave(&ctx->ctx_lock, flags); 890 /* set this inside the lock so that we can't race with aio_run_iocb() 891 * testing it and putting the iocb on the run list under the lock */ 892 if (!kiocbTryKick(iocb)) 893 run = __queue_kicked_iocb(iocb); 894 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 895 if (run) 896 aio_queue_work(ctx); 897 } 898 899 /* 900 * kick_iocb: 901 * Called typically from a wait queue callback context 902 * (aio_wake_function) to trigger a retry of the iocb. 903 * The retry is usually executed by aio workqueue 904 * threads (See aio_kick_handler). 905 */ 906 void fastcall kick_iocb(struct kiocb *iocb) 907 { 908 /* sync iocbs are easy: they can only ever be executing from a 909 * single context. */ 910 if (is_sync_kiocb(iocb)) { 911 kiocbSetKicked(iocb); 912 wake_up_process(iocb->ki_obj.tsk); 913 return; 914 } 915 916 try_queue_kicked_iocb(iocb); 917 } 918 EXPORT_SYMBOL(kick_iocb); 919 920 /* aio_complete 921 * Called when the io request on the given iocb is complete. 922 * Returns true if this is the last user of the request. The 923 * only other user of the request can be the cancellation code. 924 */ 925 int fastcall aio_complete(struct kiocb *iocb, long res, long res2) 926 { 927 struct kioctx *ctx = iocb->ki_ctx; 928 struct aio_ring_info *info; 929 struct aio_ring *ring; 930 struct io_event *event; 931 unsigned long flags; 932 unsigned long tail; 933 int ret; 934 935 /* 936 * Special case handling for sync iocbs: 937 * - events go directly into the iocb for fast handling 938 * - the sync task with the iocb in its stack holds the single iocb 939 * ref, no other paths have a way to get another ref 940 * - the sync task helpfully left a reference to itself in the iocb 941 */ 942 if (is_sync_kiocb(iocb)) { 943 BUG_ON(iocb->ki_users != 1); 944 iocb->ki_user_data = res; 945 iocb->ki_users = 0; 946 wake_up_process(iocb->ki_obj.tsk); 947 return 1; 948 } 949 950 info = &ctx->ring_info; 951 952 /* add a completion event to the ring buffer. 953 * must be done holding ctx->ctx_lock to prevent 954 * other code from messing with the tail 955 * pointer since we might be called from irq 956 * context. 957 */ 958 spin_lock_irqsave(&ctx->ctx_lock, flags); 959 960 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 961 list_del_init(&iocb->ki_run_list); 962 963 /* 964 * cancelled requests don't get events, userland was given one 965 * when the event got cancelled. 966 */ 967 if (kiocbIsCancelled(iocb)) 968 goto put_rq; 969 970 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 971 972 tail = info->tail; 973 event = aio_ring_event(info, tail, KM_IRQ0); 974 if (++tail >= info->nr) 975 tail = 0; 976 977 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 978 event->data = iocb->ki_user_data; 979 event->res = res; 980 event->res2 = res2; 981 982 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 983 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 984 res, res2); 985 986 /* after flagging the request as done, we 987 * must never even look at it again 988 */ 989 smp_wmb(); /* make event visible before updating tail */ 990 991 info->tail = tail; 992 ring->tail = tail; 993 994 put_aio_ring_event(event, KM_IRQ0); 995 kunmap_atomic(ring, KM_IRQ1); 996 997 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 998 put_rq: 999 /* everything turned out well, dispose of the aiocb. */ 1000 ret = __aio_put_req(ctx, iocb); 1001 1002 if (waitqueue_active(&ctx->wait)) 1003 wake_up(&ctx->wait); 1004 1005 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1006 return ret; 1007 } 1008 1009 /* aio_read_evt 1010 * Pull an event off of the ioctx's event ring. Returns the number of 1011 * events fetched (0 or 1 ;-) 1012 * FIXME: make this use cmpxchg. 1013 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1014 */ 1015 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1016 { 1017 struct aio_ring_info *info = &ioctx->ring_info; 1018 struct aio_ring *ring; 1019 unsigned long head; 1020 int ret = 0; 1021 1022 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1023 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1024 (unsigned long)ring->head, (unsigned long)ring->tail, 1025 (unsigned long)ring->nr); 1026 1027 if (ring->head == ring->tail) 1028 goto out; 1029 1030 spin_lock(&info->ring_lock); 1031 1032 head = ring->head % info->nr; 1033 if (head != ring->tail) { 1034 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1035 *ent = *evp; 1036 head = (head + 1) % info->nr; 1037 smp_mb(); /* finish reading the event before updatng the head */ 1038 ring->head = head; 1039 ret = 1; 1040 put_aio_ring_event(evp, KM_USER1); 1041 } 1042 spin_unlock(&info->ring_lock); 1043 1044 out: 1045 kunmap_atomic(ring, KM_USER0); 1046 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1047 (unsigned long)ring->head, (unsigned long)ring->tail); 1048 return ret; 1049 } 1050 1051 struct aio_timeout { 1052 struct timer_list timer; 1053 int timed_out; 1054 struct task_struct *p; 1055 }; 1056 1057 static void timeout_func(unsigned long data) 1058 { 1059 struct aio_timeout *to = (struct aio_timeout *)data; 1060 1061 to->timed_out = 1; 1062 wake_up_process(to->p); 1063 } 1064 1065 static inline void init_timeout(struct aio_timeout *to) 1066 { 1067 init_timer(&to->timer); 1068 to->timer.data = (unsigned long)to; 1069 to->timer.function = timeout_func; 1070 to->timed_out = 0; 1071 to->p = current; 1072 } 1073 1074 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1075 const struct timespec *ts) 1076 { 1077 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1078 if (time_after(to->timer.expires, jiffies)) 1079 add_timer(&to->timer); 1080 else 1081 to->timed_out = 1; 1082 } 1083 1084 static inline void clear_timeout(struct aio_timeout *to) 1085 { 1086 del_singleshot_timer_sync(&to->timer); 1087 } 1088 1089 static int read_events(struct kioctx *ctx, 1090 long min_nr, long nr, 1091 struct io_event __user *event, 1092 struct timespec __user *timeout) 1093 { 1094 long start_jiffies = jiffies; 1095 struct task_struct *tsk = current; 1096 DECLARE_WAITQUEUE(wait, tsk); 1097 int ret; 1098 int i = 0; 1099 struct io_event ent; 1100 struct aio_timeout to; 1101 int retry = 0; 1102 1103 /* needed to zero any padding within an entry (there shouldn't be 1104 * any, but C is fun! 1105 */ 1106 memset(&ent, 0, sizeof(ent)); 1107 retry: 1108 ret = 0; 1109 while (likely(i < nr)) { 1110 ret = aio_read_evt(ctx, &ent); 1111 if (unlikely(ret <= 0)) 1112 break; 1113 1114 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1115 ent.data, ent.obj, ent.res, ent.res2); 1116 1117 /* Could we split the check in two? */ 1118 ret = -EFAULT; 1119 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1120 dprintk("aio: lost an event due to EFAULT.\n"); 1121 break; 1122 } 1123 ret = 0; 1124 1125 /* Good, event copied to userland, update counts. */ 1126 event ++; 1127 i ++; 1128 } 1129 1130 if (min_nr <= i) 1131 return i; 1132 if (ret) 1133 return ret; 1134 1135 /* End fast path */ 1136 1137 /* racey check, but it gets redone */ 1138 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1139 retry = 1; 1140 aio_run_all_iocbs(ctx); 1141 goto retry; 1142 } 1143 1144 init_timeout(&to); 1145 if (timeout) { 1146 struct timespec ts; 1147 ret = -EFAULT; 1148 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1149 goto out; 1150 1151 set_timeout(start_jiffies, &to, &ts); 1152 } 1153 1154 while (likely(i < nr)) { 1155 add_wait_queue_exclusive(&ctx->wait, &wait); 1156 do { 1157 set_task_state(tsk, TASK_INTERRUPTIBLE); 1158 ret = aio_read_evt(ctx, &ent); 1159 if (ret) 1160 break; 1161 if (min_nr <= i) 1162 break; 1163 ret = 0; 1164 if (to.timed_out) /* Only check after read evt */ 1165 break; 1166 schedule(); 1167 if (signal_pending(tsk)) { 1168 ret = -EINTR; 1169 break; 1170 } 1171 /*ret = aio_read_evt(ctx, &ent);*/ 1172 } while (1) ; 1173 1174 set_task_state(tsk, TASK_RUNNING); 1175 remove_wait_queue(&ctx->wait, &wait); 1176 1177 if (unlikely(ret <= 0)) 1178 break; 1179 1180 ret = -EFAULT; 1181 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1182 dprintk("aio: lost an event due to EFAULT.\n"); 1183 break; 1184 } 1185 1186 /* Good, event copied to userland, update counts. */ 1187 event ++; 1188 i ++; 1189 } 1190 1191 if (timeout) 1192 clear_timeout(&to); 1193 out: 1194 return i ? i : ret; 1195 } 1196 1197 /* Take an ioctx and remove it from the list of ioctx's. Protects 1198 * against races with itself via ->dead. 1199 */ 1200 static void io_destroy(struct kioctx *ioctx) 1201 { 1202 struct mm_struct *mm = current->mm; 1203 struct kioctx **tmp; 1204 int was_dead; 1205 1206 /* delete the entry from the list is someone else hasn't already */ 1207 write_lock(&mm->ioctx_list_lock); 1208 was_dead = ioctx->dead; 1209 ioctx->dead = 1; 1210 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx; 1211 tmp = &(*tmp)->next) 1212 ; 1213 if (*tmp) 1214 *tmp = ioctx->next; 1215 write_unlock(&mm->ioctx_list_lock); 1216 1217 dprintk("aio_release(%p)\n", ioctx); 1218 if (likely(!was_dead)) 1219 put_ioctx(ioctx); /* twice for the list */ 1220 1221 aio_cancel_all(ioctx); 1222 wait_for_all_aios(ioctx); 1223 put_ioctx(ioctx); /* once for the lookup */ 1224 } 1225 1226 /* sys_io_setup: 1227 * Create an aio_context capable of receiving at least nr_events. 1228 * ctxp must not point to an aio_context that already exists, and 1229 * must be initialized to 0 prior to the call. On successful 1230 * creation of the aio_context, *ctxp is filled in with the resulting 1231 * handle. May fail with -EINVAL if *ctxp is not initialized, 1232 * if the specified nr_events exceeds internal limits. May fail 1233 * with -EAGAIN if the specified nr_events exceeds the user's limit 1234 * of available events. May fail with -ENOMEM if insufficient kernel 1235 * resources are available. May fail with -EFAULT if an invalid 1236 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1237 * implemented. 1238 */ 1239 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp) 1240 { 1241 struct kioctx *ioctx = NULL; 1242 unsigned long ctx; 1243 long ret; 1244 1245 ret = get_user(ctx, ctxp); 1246 if (unlikely(ret)) 1247 goto out; 1248 1249 ret = -EINVAL; 1250 if (unlikely(ctx || nr_events == 0)) { 1251 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1252 ctx, nr_events); 1253 goto out; 1254 } 1255 1256 ioctx = ioctx_alloc(nr_events); 1257 ret = PTR_ERR(ioctx); 1258 if (!IS_ERR(ioctx)) { 1259 ret = put_user(ioctx->user_id, ctxp); 1260 if (!ret) 1261 return 0; 1262 1263 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1264 io_destroy(ioctx); 1265 } 1266 1267 out: 1268 return ret; 1269 } 1270 1271 /* sys_io_destroy: 1272 * Destroy the aio_context specified. May cancel any outstanding 1273 * AIOs and block on completion. Will fail with -ENOSYS if not 1274 * implemented. May fail with -EFAULT if the context pointed to 1275 * is invalid. 1276 */ 1277 asmlinkage long sys_io_destroy(aio_context_t ctx) 1278 { 1279 struct kioctx *ioctx = lookup_ioctx(ctx); 1280 if (likely(NULL != ioctx)) { 1281 io_destroy(ioctx); 1282 return 0; 1283 } 1284 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1285 return -EINVAL; 1286 } 1287 1288 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1289 { 1290 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1291 1292 BUG_ON(ret <= 0); 1293 1294 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1295 ssize_t this = min((ssize_t)iov->iov_len, ret); 1296 iov->iov_base += this; 1297 iov->iov_len -= this; 1298 iocb->ki_left -= this; 1299 ret -= this; 1300 if (iov->iov_len == 0) { 1301 iocb->ki_cur_seg++; 1302 iov++; 1303 } 1304 } 1305 1306 /* the caller should not have done more io than what fit in 1307 * the remaining iovecs */ 1308 BUG_ON(ret > 0 && iocb->ki_left == 0); 1309 } 1310 1311 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1312 { 1313 struct file *file = iocb->ki_filp; 1314 struct address_space *mapping = file->f_mapping; 1315 struct inode *inode = mapping->host; 1316 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1317 unsigned long, loff_t); 1318 ssize_t ret = 0; 1319 unsigned short opcode; 1320 1321 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1322 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1323 rw_op = file->f_op->aio_read; 1324 opcode = IOCB_CMD_PREADV; 1325 } else { 1326 rw_op = file->f_op->aio_write; 1327 opcode = IOCB_CMD_PWRITEV; 1328 } 1329 1330 do { 1331 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1332 iocb->ki_nr_segs - iocb->ki_cur_seg, 1333 iocb->ki_pos); 1334 if (ret > 0) 1335 aio_advance_iovec(iocb, ret); 1336 1337 /* retry all partial writes. retry partial reads as long as its a 1338 * regular file. */ 1339 } while (ret > 0 && iocb->ki_left > 0 && 1340 (opcode == IOCB_CMD_PWRITEV || 1341 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1342 1343 /* This means we must have transferred all that we could */ 1344 /* No need to retry anymore */ 1345 if ((ret == 0) || (iocb->ki_left == 0)) 1346 ret = iocb->ki_nbytes - iocb->ki_left; 1347 1348 return ret; 1349 } 1350 1351 static ssize_t aio_fdsync(struct kiocb *iocb) 1352 { 1353 struct file *file = iocb->ki_filp; 1354 ssize_t ret = -EINVAL; 1355 1356 if (file->f_op->aio_fsync) 1357 ret = file->f_op->aio_fsync(iocb, 1); 1358 return ret; 1359 } 1360 1361 static ssize_t aio_fsync(struct kiocb *iocb) 1362 { 1363 struct file *file = iocb->ki_filp; 1364 ssize_t ret = -EINVAL; 1365 1366 if (file->f_op->aio_fsync) 1367 ret = file->f_op->aio_fsync(iocb, 0); 1368 return ret; 1369 } 1370 1371 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb) 1372 { 1373 ssize_t ret; 1374 1375 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf, 1376 kiocb->ki_nbytes, 1, 1377 &kiocb->ki_inline_vec, &kiocb->ki_iovec); 1378 if (ret < 0) 1379 goto out; 1380 1381 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1382 kiocb->ki_cur_seg = 0; 1383 /* ki_nbytes/left now reflect bytes instead of segs */ 1384 kiocb->ki_nbytes = ret; 1385 kiocb->ki_left = ret; 1386 1387 ret = 0; 1388 out: 1389 return ret; 1390 } 1391 1392 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1393 { 1394 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1395 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1396 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1397 kiocb->ki_nr_segs = 1; 1398 kiocb->ki_cur_seg = 0; 1399 return 0; 1400 } 1401 1402 /* 1403 * aio_setup_iocb: 1404 * Performs the initial checks and aio retry method 1405 * setup for the kiocb at the time of io submission. 1406 */ 1407 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1408 { 1409 struct file *file = kiocb->ki_filp; 1410 ssize_t ret = 0; 1411 1412 switch (kiocb->ki_opcode) { 1413 case IOCB_CMD_PREAD: 1414 ret = -EBADF; 1415 if (unlikely(!(file->f_mode & FMODE_READ))) 1416 break; 1417 ret = -EFAULT; 1418 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1419 kiocb->ki_left))) 1420 break; 1421 ret = security_file_permission(file, MAY_READ); 1422 if (unlikely(ret)) 1423 break; 1424 ret = aio_setup_single_vector(kiocb); 1425 if (ret) 1426 break; 1427 ret = -EINVAL; 1428 if (file->f_op->aio_read) 1429 kiocb->ki_retry = aio_rw_vect_retry; 1430 break; 1431 case IOCB_CMD_PWRITE: 1432 ret = -EBADF; 1433 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1434 break; 1435 ret = -EFAULT; 1436 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1437 kiocb->ki_left))) 1438 break; 1439 ret = security_file_permission(file, MAY_WRITE); 1440 if (unlikely(ret)) 1441 break; 1442 ret = aio_setup_single_vector(kiocb); 1443 if (ret) 1444 break; 1445 ret = -EINVAL; 1446 if (file->f_op->aio_write) 1447 kiocb->ki_retry = aio_rw_vect_retry; 1448 break; 1449 case IOCB_CMD_PREADV: 1450 ret = -EBADF; 1451 if (unlikely(!(file->f_mode & FMODE_READ))) 1452 break; 1453 ret = security_file_permission(file, MAY_READ); 1454 if (unlikely(ret)) 1455 break; 1456 ret = aio_setup_vectored_rw(READ, kiocb); 1457 if (ret) 1458 break; 1459 ret = -EINVAL; 1460 if (file->f_op->aio_read) 1461 kiocb->ki_retry = aio_rw_vect_retry; 1462 break; 1463 case IOCB_CMD_PWRITEV: 1464 ret = -EBADF; 1465 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1466 break; 1467 ret = security_file_permission(file, MAY_WRITE); 1468 if (unlikely(ret)) 1469 break; 1470 ret = aio_setup_vectored_rw(WRITE, kiocb); 1471 if (ret) 1472 break; 1473 ret = -EINVAL; 1474 if (file->f_op->aio_write) 1475 kiocb->ki_retry = aio_rw_vect_retry; 1476 break; 1477 case IOCB_CMD_FDSYNC: 1478 ret = -EINVAL; 1479 if (file->f_op->aio_fsync) 1480 kiocb->ki_retry = aio_fdsync; 1481 break; 1482 case IOCB_CMD_FSYNC: 1483 ret = -EINVAL; 1484 if (file->f_op->aio_fsync) 1485 kiocb->ki_retry = aio_fsync; 1486 break; 1487 default: 1488 dprintk("EINVAL: io_submit: no operation provided\n"); 1489 ret = -EINVAL; 1490 } 1491 1492 if (!kiocb->ki_retry) 1493 return ret; 1494 1495 return 0; 1496 } 1497 1498 /* 1499 * aio_wake_function: 1500 * wait queue callback function for aio notification, 1501 * Simply triggers a retry of the operation via kick_iocb. 1502 * 1503 * This callback is specified in the wait queue entry in 1504 * a kiocb (current->io_wait points to this wait queue 1505 * entry when an aio operation executes; it is used 1506 * instead of a synchronous wait when an i/o blocking 1507 * condition is encountered during aio). 1508 * 1509 * Note: 1510 * This routine is executed with the wait queue lock held. 1511 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1512 * the ioctx lock inside the wait queue lock. This is safe 1513 * because this callback isn't used for wait queues which 1514 * are nested inside ioctx lock (i.e. ctx->wait) 1515 */ 1516 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1517 int sync, void *key) 1518 { 1519 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1520 1521 list_del_init(&wait->task_list); 1522 kick_iocb(iocb); 1523 return 1; 1524 } 1525 1526 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1527 struct iocb *iocb) 1528 { 1529 struct kiocb *req; 1530 struct file *file; 1531 ssize_t ret; 1532 1533 /* enforce forwards compatibility on users */ 1534 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 || 1535 iocb->aio_reserved3)) { 1536 pr_debug("EINVAL: io_submit: reserve field set\n"); 1537 return -EINVAL; 1538 } 1539 1540 /* prevent overflows */ 1541 if (unlikely( 1542 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1543 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1544 ((ssize_t)iocb->aio_nbytes < 0) 1545 )) { 1546 pr_debug("EINVAL: io_submit: overflow check\n"); 1547 return -EINVAL; 1548 } 1549 1550 file = fget(iocb->aio_fildes); 1551 if (unlikely(!file)) 1552 return -EBADF; 1553 1554 req = aio_get_req(ctx); /* returns with 2 references to req */ 1555 if (unlikely(!req)) { 1556 fput(file); 1557 return -EAGAIN; 1558 } 1559 1560 req->ki_filp = file; 1561 ret = put_user(req->ki_key, &user_iocb->aio_key); 1562 if (unlikely(ret)) { 1563 dprintk("EFAULT: aio_key\n"); 1564 goto out_put_req; 1565 } 1566 1567 req->ki_obj.user = user_iocb; 1568 req->ki_user_data = iocb->aio_data; 1569 req->ki_pos = iocb->aio_offset; 1570 1571 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1572 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1573 req->ki_opcode = iocb->aio_lio_opcode; 1574 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1575 INIT_LIST_HEAD(&req->ki_wait.task_list); 1576 1577 ret = aio_setup_iocb(req); 1578 1579 if (ret) 1580 goto out_put_req; 1581 1582 spin_lock_irq(&ctx->ctx_lock); 1583 aio_run_iocb(req); 1584 if (!list_empty(&ctx->run_list)) { 1585 /* drain the run list */ 1586 while (__aio_run_iocbs(ctx)) 1587 ; 1588 } 1589 spin_unlock_irq(&ctx->ctx_lock); 1590 aio_put_req(req); /* drop extra ref to req */ 1591 return 0; 1592 1593 out_put_req: 1594 aio_put_req(req); /* drop extra ref to req */ 1595 aio_put_req(req); /* drop i/o ref to req */ 1596 return ret; 1597 } 1598 1599 /* sys_io_submit: 1600 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1601 * the number of iocbs queued. May return -EINVAL if the aio_context 1602 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1603 * *iocbpp[0] is not properly initialized, if the operation specified 1604 * is invalid for the file descriptor in the iocb. May fail with 1605 * -EFAULT if any of the data structures point to invalid data. May 1606 * fail with -EBADF if the file descriptor specified in the first 1607 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1608 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1609 * fail with -ENOSYS if not implemented. 1610 */ 1611 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr, 1612 struct iocb __user * __user *iocbpp) 1613 { 1614 struct kioctx *ctx; 1615 long ret = 0; 1616 int i; 1617 1618 if (unlikely(nr < 0)) 1619 return -EINVAL; 1620 1621 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1622 return -EFAULT; 1623 1624 ctx = lookup_ioctx(ctx_id); 1625 if (unlikely(!ctx)) { 1626 pr_debug("EINVAL: io_submit: invalid context id\n"); 1627 return -EINVAL; 1628 } 1629 1630 /* 1631 * AKPM: should this return a partial result if some of the IOs were 1632 * successfully submitted? 1633 */ 1634 for (i=0; i<nr; i++) { 1635 struct iocb __user *user_iocb; 1636 struct iocb tmp; 1637 1638 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1639 ret = -EFAULT; 1640 break; 1641 } 1642 1643 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1644 ret = -EFAULT; 1645 break; 1646 } 1647 1648 ret = io_submit_one(ctx, user_iocb, &tmp); 1649 if (ret) 1650 break; 1651 } 1652 1653 put_ioctx(ctx); 1654 return i ? i : ret; 1655 } 1656 1657 /* lookup_kiocb 1658 * Finds a given iocb for cancellation. 1659 */ 1660 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1661 u32 key) 1662 { 1663 struct list_head *pos; 1664 1665 assert_spin_locked(&ctx->ctx_lock); 1666 1667 /* TODO: use a hash or array, this sucks. */ 1668 list_for_each(pos, &ctx->active_reqs) { 1669 struct kiocb *kiocb = list_kiocb(pos); 1670 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1671 return kiocb; 1672 } 1673 return NULL; 1674 } 1675 1676 /* sys_io_cancel: 1677 * Attempts to cancel an iocb previously passed to io_submit. If 1678 * the operation is successfully cancelled, the resulting event is 1679 * copied into the memory pointed to by result without being placed 1680 * into the completion queue and 0 is returned. May fail with 1681 * -EFAULT if any of the data structures pointed to are invalid. 1682 * May fail with -EINVAL if aio_context specified by ctx_id is 1683 * invalid. May fail with -EAGAIN if the iocb specified was not 1684 * cancelled. Will fail with -ENOSYS if not implemented. 1685 */ 1686 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb, 1687 struct io_event __user *result) 1688 { 1689 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1690 struct kioctx *ctx; 1691 struct kiocb *kiocb; 1692 u32 key; 1693 int ret; 1694 1695 ret = get_user(key, &iocb->aio_key); 1696 if (unlikely(ret)) 1697 return -EFAULT; 1698 1699 ctx = lookup_ioctx(ctx_id); 1700 if (unlikely(!ctx)) 1701 return -EINVAL; 1702 1703 spin_lock_irq(&ctx->ctx_lock); 1704 ret = -EAGAIN; 1705 kiocb = lookup_kiocb(ctx, iocb, key); 1706 if (kiocb && kiocb->ki_cancel) { 1707 cancel = kiocb->ki_cancel; 1708 kiocb->ki_users ++; 1709 kiocbSetCancelled(kiocb); 1710 } else 1711 cancel = NULL; 1712 spin_unlock_irq(&ctx->ctx_lock); 1713 1714 if (NULL != cancel) { 1715 struct io_event tmp; 1716 pr_debug("calling cancel\n"); 1717 memset(&tmp, 0, sizeof(tmp)); 1718 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1719 tmp.data = kiocb->ki_user_data; 1720 ret = cancel(kiocb, &tmp); 1721 if (!ret) { 1722 /* Cancellation succeeded -- copy the result 1723 * into the user's buffer. 1724 */ 1725 if (copy_to_user(result, &tmp, sizeof(tmp))) 1726 ret = -EFAULT; 1727 } 1728 } else 1729 ret = -EINVAL; 1730 1731 put_ioctx(ctx); 1732 1733 return ret; 1734 } 1735 1736 /* io_getevents: 1737 * Attempts to read at least min_nr events and up to nr events from 1738 * the completion queue for the aio_context specified by ctx_id. May 1739 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1740 * if nr is out of range, if when is out of range. May fail with 1741 * -EFAULT if any of the memory specified to is invalid. May return 1742 * 0 or < min_nr if no events are available and the timeout specified 1743 * by when has elapsed, where when == NULL specifies an infinite 1744 * timeout. Note that the timeout pointed to by when is relative and 1745 * will be updated if not NULL and the operation blocks. Will fail 1746 * with -ENOSYS if not implemented. 1747 */ 1748 asmlinkage long sys_io_getevents(aio_context_t ctx_id, 1749 long min_nr, 1750 long nr, 1751 struct io_event __user *events, 1752 struct timespec __user *timeout) 1753 { 1754 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1755 long ret = -EINVAL; 1756 1757 if (likely(ioctx)) { 1758 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1759 ret = read_events(ioctx, min_nr, nr, events, timeout); 1760 put_ioctx(ioctx); 1761 } 1762 1763 return ret; 1764 } 1765 1766 __initcall(aio_setup); 1767 1768 EXPORT_SYMBOL(aio_complete); 1769 EXPORT_SYMBOL(aio_put_req); 1770 EXPORT_SYMBOL(wait_on_sync_kiocb); 1771